AI Revolutionizes Seismic Vulnerability Assessment

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Jul 3 2024

In an article published in the journal Applied Sciences, researchers presented a method to improve seismic vulnerability assessment using machine learning (ML) models within rapid visual screening (RVS) frameworks.

By training binary classifiers and applying ML feature attribution techniques, the authors demonstrated that ML-based models could outperform traditional engineering practices. The paper proposed a linearization of ML models, inspired by established codes, to recalibrate RVS procedures for improved accuracy in ranking structures by their seismic vulnerability.

Background

The seismic vulnerability of structures is a critical engineering challenge, especially given the rapid increase in global urbanization and the presence of older buildings not sticking to modern safety standards. RVS procedures, developed by entities such as the Federal Emergency Management Agency (FEMA), New Zealand Society for Earthquake Engineering, and others, provide an initial assessment of buildings' vulnerability using expert-defined structural features and weights.

Previous approaches have varied across countries, focusing on different structural characteristics and weights, with accuracy and efficiency validated against near-field earthquake data. Recent advancements in ML have introduced potential enhancements to RVS methodologies. Prior research has explored using ML for tasks like identifying structural features from images and predicting damage degrees.

The present paper introduced a novel application of ML classification models for seismic vulnerability ranking, demonstrating their superiority over traditional methods. Additionally, it proposed a linearization of ML models using feature attribution techniques to recalibrate RVS procedures, thus bridging the gap between ML advancements and existing engineering practices.

Data Collection and Analytical Methods

The researchers used a dataset of 457 structures from the 1999 Athens earthquake, encompassing a wide range of damage states from minor to total collapse. To mitigate local effects, the dataset included structures from various regions in the Athens Metropolitan area. The damage states were classified into four categories: Black (severe damage), Red (extensive damage), Yellow (moderate damage), and Green (minor damage). Various seismic codes (FEMA, Canada, India, New Zealand, Greek Organization for Seismic Planning and Protection (OASP)) considered distinct structural features and assigned different weights to compute a vulnerability index.

The research aimed to improve RVS procedures using ML models. A binary classification model was developed to rank pairs of structures based on their seismic vulnerability. The gradient boosting (GB) classifier, an ensemble method combining multiple decision trees, was used due to its superior performance in previous studies. Data preprocessing steps included addressing class imbalance through undersampling and transforming features into numerical values.

An innovative recalibration approach was proposed, leveraging shapley additive explanations (SHAP) values to assign weights to structural features, aligning with the additive nature of traditional seismic codes. SHAP values, derived from cooperative game theory, decomposed the model's predictions into contributions from individual features. These values were aggregated to form a single weight vector for each seismic code, improving the RVS procedures' accuracy.

The proposed methodology showed significant performance improvements over traditional RVS approaches. The researchers highlighted the potential for applying this ML-driven recalibration process to larger datasets and diverse seismic conditions, paving the way for more accurate and tailored seismic vulnerability assessments globally.

Findings and Analysis

Results indicated that the OASP 2000 and 2004 seismic codes outperformed others using traditional RVS. However, ML models significantly improved accuracy across all damage categories, especially for severe damage (Black and Red classes). For instance, the FEMA 2002 code's accuracy for the Black class increased from 40% to 61% using ML. The Indian, Canadian, and New Zealand codes also saw over a 10% improvement for the Black class. OASP showed a 5% improvement, maintaining the highest initial performance.

The proposed ML-based recalibration further improved accuracy for most seismic codes, particularly for severe damage categories. However, the Canadian Code, which used a fundamentally different scoring method, showed minimal improvement with this linear recalibration approach.

Overall, the study highlighted the benefits of using ML for seismic vulnerability ranking, demonstrating that ML models can learn complex, non-linear relationships, significantly outperforming traditional engineering practices and improving the accuracy of seismic damage predictions.

Conclusion

In conclusion, the researchers showcased the potential of ML in enhancing seismic vulnerability assessment. By leveraging ML models and innovative recalibration techniques inspired by established seismic codes, significant improvements in accuracy were achieved across various damage categories.

The findings underscored ML's ability to learn complex patterns and outperform traditional engineering practices, offering a promising pathway for more precise and reliable seismic risk assessments globally. Future research could explore broader datasets and diverse structural types to further refine and expand the applicability of ML-driven methodologies in seismic engineering.